EY Americas' AI Leaders on Optimizing Your AI Workforce

Traci Gusher, EY Americas data and AI leader, and David Guarrera, its generative AI leader, spoke to AI Business Editor Deborah Yao about best practices for assembling an AI workforce. They say it is not as expensive or difficult as one might think since for most companies there is no need to hire elite engineers. Also, they believe prompt engineering will not last as a solo career in the future.

Listen to the podcast below or read the edited transcript.

AI Business: Today, we want to talk about how AI is changing the structure and nature of the workforce. EY recently released its emerging tech at work survey and we're going to do a deep dive into its findings. First of all, can you tell me what are the main takeaways from the survey?

Traci Gusher: There are a few things that I think popped out for me out of the data. One is that respondents don't believe their senior executives are investing enough, or deploying AI fast enough into their organizations. Among employees it was about 60% who believe so, and manager-level it was about 64%. It was just very clear that the perception is that senior leaders and organizations really don't have the urgency that the employees feel is necessary.

When we asked about the barriers to adoption, two key ones popped out for me. One was that the respondents in this survey believed that one of the top barriers was ethical or moral concerns. … In addition to that, skills gaps and shortages of labor to deploy artificial intelligence at scale is also a key point.

AI Business: What are the implications of this gap between what workers want and what top management believes is best for the company?

Gusher: The labor force that we have today is very eager to be capitalizing on the use of emerging technologies. … We (quickly) became accustomed to leveraging ChatGPT to help us in a variety of ways (because) it was so easy. And now these employees are sitting in front of their computers doing their work, and they're saying, ‘Why is technology so hard for my company, when it's become so easy for me, in my personal life?’

The companies that are able to move faster and actually get more technology with ease into the hands of their employees, they're going to end up getting the top talent, because that's where this talent is going to want to be.

David Guarrera: Part of it is the gap between what junior employees know is useful and senior employees who are maybe a little bit more risk-averse. It's a tale as old as time. … I think as you get more senior there's a natural skepticism and a fear of change because you rose (up the ranks) in a previous system.

Nevertheless, I do think that with this technology, a little bit of skepticism is warranted. And it's worth thinking through some of the risks of these technologies. … You very quickly realize that this technology is going to change the world and that we really need to get in front of it.

AI Business: So how do you get workers and management to agree?

Gusher: Some of this is just a communication gap. The perception of employees is that senior executives are not moving fast enough and that they're not investing. But a lot of the organizations that we're working with are investing significantly in artificial intelligence, but it isn't something that gets deployed and rolled out overnight. So more communication is one way to bridge the gap of understanding and actuality as well as what are feasible and reasonable timelines to get this kind of tech into the hands of employees and really impacting the processes and their daily tasks.

AI Business: There are several well-documented risks of generative AI, like hallucinations, privacy violations because they train on people's proprietary data and others. When you're talking to clients, what are some risks they believe are more pressing than others?

Guarrera: First and foremost, hallucinations - the fact that these models not infrequently make up answers and are very confident about those answers. There have been lots of concern about privacy issues as well. There have been concerns about IP infringement and cybersecurity concerns. What we're seeing is that a lot of the first wave of how people are deploying this technology is in the back office, to really act as a copilot to help people do their own work faster.

A lot of people say, ‘how could you possibly use these models? They make things up all that time.’ To which I would say, have you ever worked with a junior employee? They get things wrong all the time. And the reason organizations work at all is because embedded are systems of oversight and review. … That can go a long way in terms of mitigating the risks of hallucination and bias.

AI Business: Do companies really have the personnel to manage this new technology? And if you need to assemble a team, what would that team look like?

Guarrera: This issue of having a skilled workforce and upskilling is a very interesting one because there's sort of a premise out there that we're going to need very highly skilled people to join every organization. That is somehow both true and not true at the same time. It's definitely true that to build large scale, deployable, custom machine learning generative AI, large language model-based systems, you're going to need very, very talented machine learning engineers or generative AI engineers. This problem is one that has existed for a while for anybody looking to deploy machine learning or AI.

But entering into the golden age of AI, people are going to find themselves making the decision to upskill their workforce or to contract that out a lot more often. That's one component. There's another sense, though, for which the new advancements in generative AI actually solve some of the problems of not having an overly skilled workforce. One premise is that they can effectively democratize complex processes or complex analyses by making the interface through normal speech and natural language.

And so for AI processes that don't need to necessarily be customized or be particularly detailed, all of a sudden, people without necessarily deep AI skills, have access to these tools. You can now, for example, upload some complex data to ChatGPT and just by talking to it in natural language, deploy complex algorithms that previously were really the domain of those who knew MLOps or knew how to code in Python. And so all of a sudden, something you might have paid somebody else to do in the past, either by hiring them, or contracting it out, is now at your fingertips.

We're entering an interesting era of no-code or low-code solutions, where the large language model becomes an interface for all sorts of complex tools that have existed before but really required a lot of domain expertise to unlock.

AI Business: What roles would you require on an AI team?

Gusher: By and large, the same skillsets that are needed for managing generative AI are the same types of roles that are going to be needed for generative AI - you still need data scientists, you still need data and software engineers, you need folks that are skilled in model management and MLOps. David, maybe you have some other thoughts in terms of additional skillsets above and beyond that, perhaps in prompt engineering or the like.

Guarrera: There is a new and emerging field of so-called prompt engineering, which is having the skill to talk to these models the right way. And if you've played with these models at all, you know that how you ask your questions, or how you phrase your prompts, can make a huge difference in terms of model performance and what you get back − it's a new paradigm for us. It can sometimes feel like you're querying the Oracle of Delphi, and sometimes you get back exactly what you want, and sometimes the model decides that to heck with you, I'm not going to listen to some parts of your prompts. The ability to get the best out of large language models is a skill, and it's one that can be taught, and it's one that's valuable - separate and apart from traditional data skills.

Gusher: I think prompt engineering is a skillset that should be learned by everyone. … It should be no different than other new skills that employees get trained on. I'm not of the opinion that it warrants an actual full time, job or role.

Guarrera: I agree. There's been a lot of interesting evidence recently that we're moving towards a place where prompt engineering is going to actually be handled by another generative AI model. I'll give you an example. I have been using generative image models a lot for making sides and presentations. There's a large language model that I use that is specifically engineered to get great prompts at an image model. So I am not a great prompt engineer, but I'm able to go to this tool, give a brief description of what I want and out of this initial prompt engineering large language model comes a really detailed prompt that produces beautiful images.

There's been a lot of arguments about whether or not prompt engineering is going to be a career or not. Some of the current thinking suggests that in two or three years, there will certainly be a basic generative AI prompting skillset similar to the skillset of all digital natives after the internet came along. But a lot of the complex approaches to prompting are going to be subsumed by other models that can automatically do this.

AI Business: How about for AI experts? We've seen headlines that say they make a million dollars a year. Is that real? Or is that just for the really gifted AI practitioners?

Gusher: Our average AI data scientist isn't making a million dollars a year because it would be pretty extraordinary. Certainly AI engineers, AI data scientists are a skilled and well-paid workforce. But I think these million dollar tags are a bit of hype.

Guarrera: If that's the case, Traci and myself both need to be asking for raises. You look at one very senior job listing at a place like Netflix and then write a news article that everybody's getting paid a million dollars. Not everybody's getting paid a million dollars. Nevertheless, this workforce is extremely skilled and their skills are in high demand right now − and that's only going to grow.

AI Business: There are fears that AI will replace human jobs. The hype is that it's assistive to humans and won't replace human jobs. But how can it not? And that's what actually makes it really attractive to businesses is they can cut employee costs. Machines are more reliable and they don't get tired. What's your take on that?

Guarrera: First of all, I should caveat my answer by saying that I am not a labor economist. Anything I'm saying is a bit of speculation, although that's never stopped me before. There's a line going around that it's not AI that's going to take your job. It's somebody that knows AI that's going to take your job.

You can think back to other massive efficiency gains that we've been through before during the rise of the internet and the web. All of a sudden, people can be radically more efficient than they could be before. Did organizations change their workforce? Did EY or Goldman Sachs decrease their incoming (new employee) class sizes? No, I don't think they did. They just got a lot more done and added a lot more value to the organization and to the economy.

I think something similar will happen here. I do not think robots are going to take all of our jobs. And on the whole, this may even be positive in terms of workforce growth numbers. However, to your point, anytime, anything like this happens the truth is that there are always winners and losers. The exact makeup of the workforce and the skills necessary are definitely going to change. Going back to the internet example, that has made us all more productive. But if you were an archivist who worked in the basement of an organization, that is a job that's probably not there anymore. I think we will certainly see a reshuffling.

Gusher: The fact is that there will be jobs that are replaced by AI - there just will be. Just like there have been jobs replaced by a lot of other technology and innovation. But ultimately, I do believe that the balance of need is equal or greater − where there are jobs being eliminated, there are jobs that are going to be created. Will they require the same types of skills and resources? Probably not. … Some folks are going to come out on top and be in more demand, and others are going to have to rescale and do something different.

There are jobs that are going to be replaced, there are jobs that are going to be augmented, and then those that are going to increase. The augmentation is where it is going to touch almost every role, right? So in some way, shape or form, your job will be augmented by this technology. But will your job go away? That's a far more limited spectrum of roles, I think.

AI Business: What will the future workforce look like with AI embedded in it?

Guarrera: Roles like data scientists and AI engineers are going to grow in importance and they're going to continue to be a valuable commodity. I do also want to reiterate that I don't think that means everybody working in an organization will need to be a deep AI expert. However, everybody in the workforce will need to naturally move to a point of AI literacy.

Going back to the rise of the internet and the web, everybody in the workforce now has a minimum amount of digital and online literacy. … I think it will be the same way with generative AI and with AI. We'll look back with quaintness on these early times where we're all talking about specific use cases, or talking about how to best prompt the model, because I think this is going to be a basic part of the skills that everybody in the workforce learns, and will be part our upbringing, training and education.

AI Business: So you're saying that we'll all be AI practitioners?

Guarrera: Yes. I've been in and around this field for 10 or 15 years now and I've never seen anything like this happen before. Every day, every week, there are more fundamental gains, and all of a sudden, models are doing things that we never thought possible. It is an exciting time to be alive. It is an exciting time to be working in this field and really just holding on and going along for the ride.

Gusher: We had a wave of AI seven or eight years ago, when a lot of organizations started really investing in artificial intelligence, MVPs (minimum viable products) and production models. That wave kind of slowed a little bit as folks realized AI is not easy − the training of it and the production. They realized it is hard.

But now we spiked again. I actually think that unlike the last wave we saw, this spike is going to perpetuate for a very long time because of the advancements we have had in how we can accelerate to value using this and other AI. I think I would close with, ‘the hype is real.’ I am as excited as anybody else to see just how much we can do with it.